Field-based hyperspectral imaging for detection and spatial mapping of fusarium head blight in wheat

Abstract

Fusarium head blight (FHB) poses a substantial threat to cereal crop production, significantly affecting both grain yield and quality by producing harmful mycotoxins such as deoxynivalenol (DON), which is detrimental to human and animal health. To manage this threat effectively, precise detection and mapping of FHB spatial distribution at the field level are crucial. This study aimed to detect and map FHB in four commercial winter wheat fields in Belgium and Lithuania using a push-broom hyperspectral camera (400–1000 nm), mounted on a tractor. The on-line collected hyperspectral data were first subjected to a linear regression model to segment wheat ears from the background using a linear regression model, achieving a precision of 0.99. The segmented hyperspectral data were then correlated with FHB severity, assessed by means of groundtruth captured RGB images using two dataset. The first dataset (M1) combined data from both countries, whereas the second dataset (M2) used data from the three fields in Lithuania only. The two datasets were then subjected to four machine learning (ML) modelling techniques, namely, extra trees regression (ETR), random forest regression (RFR), support vector regression (SVR), and one-dimensional convolutional neural network (1DCNN). Once validated using an independent validation set, these models were used to predict and map FHB using the on-line collected spectra in the four fields. Additionally, recursive feature elimination (RFE) and mutual information (MI) approaches to select the optimal wavebands for FHB detection were employed. Results demonstrated the capability of ETR to predict FHB severity successfully, surpassing the other ML models, achieving coefficients of determination (R²) values of 0.68 and 0.79 for M1 and M2, respectively. The residual prediction deviation (RPD) values recorded were 1.77 for M1 and 2.18 for M2, and the ration of performance to inter-quartile range (RPIQ) values were 2.89 and 3.51, respectively. Moreover, M2 showed enhanced model accuracy for the used ML models, except for SVM. The application of MI on ETR significantly improved the predictive accuracy, with R² values of 0.75 for M1 and 0.82 for M2, In contrast, the application of RFE did not result in any improvement in the models effectiveness, as evidenced by R² values of 0.65 and 0.75 for M1 and M2, respectively. A comparison between the predicted points from the on-line scanning and ground truth maps shows varying levels of spatial similarity with a kappa value reaching 0.58. These results confirm the potential of integrating hyperspectral imaging with ML models for effective detection and spatial mapping of FHB in wheat fields.